Unlock free Digest editor
Roula Khalaf, editor of FT, chooses her favorite story in this weekly newsletter.
Openai launched its largest artificial intelligence model with less so -called “hallucinations”, the latest in the rapture of new editions of AI by US technological groups.
The San Francisco-based company discovered GPT-4.5 on Thursday, its long-awaited updating technology that supports its popular Chatgpt. In early tests, the hallucination rate, where AI systems create inaccurate information, was 37 percent compared to almost 60 percent on its predecessor GPT-4O.
With Gpt-4.5, Open It continues to bet on large, expensive large-language models, despite the appearance of very capable smaller products, such as Chinese start-up Deepseek R1, which are open, cheaper and more affordable to develop.
It comes as competition is growing in a rapidly developed AI industry, as technological groups have rushed to launch their latest models in recent weeks. Anthropic discovered his Claude 3.7 sonnet on Monday, which was last week after the launch of the latest Xai model, Elona Male, Grok 3.
Openii said on Thursday that GPT-4.5 had “a wider knowledge and a deeper understanding of the world, which led to reduced hallucinations and greater reliability in a wide range of topics.”
“With every new order of calculation, new opportunities come,” the company said, adding that GPT-4.5 “is on the border of what is possible in unbeatable learning.”
Openii was headed by a global AI industry race, collecting dozens of billions from investors to finance larger models with increased capabilities that require huge amounts of computer forces.
Currently, in conversations with SoftBank and other investors to raise up to $ 40 billion on value estimate $ 300 billionIncluding new money, while Anthropic also raises funding around $ 3.5 billion at $ 60 billion plus, according to two people with knowledge of the process.
However, huge costs of launching larger models have cited the Openi to consider the withdrawal of the GPT-4.5 approach developers in the midst of the fierce competition of rival prices.
Openii said the GPT-4.5 review will be available to the developers who pay to use Openi’s models through the application programming interface (API), this approach could be abolished in the future.
AI groups largely generate income through the paid approach to API and individual subscriptions. Openii said he would see how developers use a powerful model and whether it is worth offered given the high starting costs.
“GPT-4.5 is a very large and computer intense model, making it more expensive than and not a replacement for the GPT-4o [its predecessor]. Because of this, we evaluate whether we continue to serve it in the long-term API because we are balanced by supporting the current opportunities with the construction of future models, “the company said.
OPENAI CEO Sam Altman said earlier that GPT-4 would cost more than $ 100 million to train, and such costs are widely expected to increase as the size and capabilities of the model and require more computer power for training and starting.
Additional reporting George Hammond in San Francisco and Melissa Heikkilä in London
